Re: [RFC PATCH 5/5] mm proc/task_mmu.c: add hugetlb specific routine for clear_refs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Feb 10, 2021 at 04:03:22PM -0800, Mike Kravetz wrote:
> There was is no hugetlb specific routine for clearing soft dirty and
> other referrences.  The 'default' routines would only clear the
> VM_SOFTDIRTY flag in the vma.
> 
> Add new routine specifically for hugetlb vmas.
> 
> Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> ---
>  fs/proc/task_mmu.c | 110 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 110 insertions(+)
> 
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index 829b35016aaa..f06cf9b131a8 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -1116,6 +1116,115 @@ static inline void clear_soft_dirty_pmd(struct vm_area_struct *vma,
>  }
>  #endif
>  
> +#ifdef CONFIG_HUGETLB_PAGE
> +static inline bool huge_pte_is_pinned(struct vm_area_struct *vma,
> +					unsigned long addr, pte_t pte)
> +{
> +	struct page *page;
> +
> +	if (likely(!atomic_read(&vma->vm_mm->has_pinned)))
> +		return false;
> +	page = pte_page(pte);
> +	if (!page)
> +		return false;
> +	return page_maybe_dma_pinned(page);
> +}
> +
> +static int clear_refs_hugetlb_range(pte_t *ptep, unsigned long hmask,
> +				unsigned long addr, unsigned long end,
> +				struct mm_walk *walk)
> +{
> +	struct clear_refs_private *cp = walk->private;
> +	struct vm_area_struct *vma = walk->vma;
> +	struct hstate *h = hstate_vma(walk->vma);
> +	unsigned long adj_start = addr, adj_end = end;
> +	spinlock_t *ptl;
> +	pte_t old_pte, pte;
> +
> +	/*
> +	 * clear_refs should only operate on complete vmas.  Therefore,
> +	 * values passed here should be huge page aligned and huge page
> +	 * size in length.  Quick validation before taking any action in
> +	 * case upstream code is changed.
> +	 */
> +	if ((addr & hmask) != addr || end - addr != huge_page_size(h)) {
> +		WARN_ONCE(1, "%s passed unaligned address\n", __func__);
> +		return 1;
> +	}

I wouldn't worry too much on the interface change - The one who will change the
interface should guarantee all existing hooks will still work, isn't it? :)

It's slightly confusing to me on why "clear_refs should only operate on
complete vmas" is related to the check, though.

> +
> +	ptl = huge_pte_lock(hstate_vma(vma), walk->mm, ptep);
> +
> +	/* Soft dirty and pmd sharing do not mix */

Right, this seems to be a placeholder for unsharing code.

Though maybe we can do that earlier in pre_vma() hook?  That should be per-vma
rather than handling one specific huge page here, hence more efficient imho.

this reminded me that I should also better move hugetlb_unshare_all_pmds() of
my other patch into hugetlb.c, so that this code can call it.  Currently it's a
static function in userfaultfd.c.

> +
> +	pte = huge_ptep_get(ptep);
> +	if (!pte_present(pte))
> +		goto out;
> +	if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
> +		goto out;
> +
> +	if (cp->type == CLEAR_REFS_SOFT_DIRTY) {

Maybe move this check into clear_refs_test_walk()?  We can bail out earlier if:

      (is_vm_hugetlb_page(vma) && (type != CLEAR_REFS_SOFT_DIRTY))

> +		if (huge_pte_is_pinned(vma, addr, pte))
> +			goto out;

Out of topic of this patchset, but it's definitely a pity that we can't track
soft dirty for pinned pages.  Currently the assumption of the pte code path is:
"if this page can be DMA written then we won't know whether data changed after
all, then tracking dirty is meaningless", however that's prone to change when
new hardwares coming, say, IOMMU could start to trap DMA writes already.

But again that's another story.. and we should just follow what we do with
non-hugetlbfs for sure here, until some day if we'd like to revive soft dirty
tracking with pinned pages.

> +
> +		/*
> +		 * soft dirty and pmd sharing do not work together as
> +		 * per-process is tracked in ptes, and pmd sharing allows
> +		 * processed to share ptes.  We unshare any pmds here.
> +		 */
> +		adjust_range_if_pmd_sharing_possible(vma, &adj_start, &adj_end);

Ideally when reach here, huge pmd sharing won't ever exist, right?  Then do we
still need to adjust the range at all?

Thanks,

-- 
Peter Xu




[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Index of Archives]     [Kernel Development]     [Kernel Newbies]     [IDE]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite Info]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux ATA RAID]     [Samba]     [Linux Media]     [Device Mapper]

  Powered by Linux